Data Science for Material Property Interpretation: Machine Learning with Microscopy Data
Program Organizers: Alex Belianinov, Oak Ridge National Laboratory; Ichiro Takeuchi, University of Maryland; Jeff Simmons, Wright Patterson Air Force Research Laboratory; Jason Hattrick-Simpers, National Institute of Standards and Technology
Monday 8:00 AM
September 30, 2019
Location: Oregon Convention Center
Session Chair: Alex Belianinov, Oak Ridge National Laboratory
8:00 AM Invited
Data Science and the MGI: James Warren1; 1National Institute of Standards and Technology
The National Institute of Standards and Technology has focused its support of the US Materials Genome Initiative on the management, quality, and use of data to accelerated the discovery, design, development, and deployment of new materials. While the program foresaw the day when "data-driven" materials science would be a rising paradigm for research, the recent demonstrated successes, in a variety of disciplines, of deep learning and other advanced artificial intelligence algorithms has changed the conversation on the value of AI while simultaneously reinvigorating interest in related techniques. This confluence has left the MGI perfectly poised to capitalize on these remarkable developments. I will examine the progress we have made in getting data "AI-ready" and how this dovetails with our development of materials models, and the concomitant challenges in interpretation.
Workflows for Curation and Analysis of Microstructure-Aware Materials Data: Application to Aging of U-Nb Alloys: Robert Hackenberg1; Logan Ward2; 1Los Alamos National Lab; 2University of Chicago
Considering multiple data sets can mitigate the scarcity of data available for understanding and predict aging behavior in a given material. The challenge then becomes capturing heterogeneous data types (e.g., micrographs, diffraction patterns, etc.) from many different experiments and data sources – a process that requires significant, careful data curation and can be very time consuming. In this work, we explore how one can use modern data collection, storage, and analysis technologies to accelerate the development of aging models. In particular, we have developed software to store hardness data collected for uranium-niobium alloys over 70 years of research, and use automated data analysis tools to create models to predict how the hardness will change over time-at-temperature. In addition to providing the tools to accelerate model development, we also implemented methods for storing the rationale behind decisions made during the modeling process in order to enable better reproducibility.
Data Analytics for Correlative Multimodal Chemical and Functional Imaging: Anton Ievlev1; Olga Ovchinnikova1; 1Oak Ridge National Laboratory
Advancing of functional materials requires understanding and control their structure, chemistry and function on the nanoscale. While much of the chemical and structural properties can be studied on macro-scale systems, there is a lack of information about chemical properties on the nanoscale and its correlation to the structure. Here we suggest multimodal chemical and functional imaging approach combining scanning probe microscopy with mass spectrometry and optical spectroscopy to unravel behavior of functional materials. This approach transcends existing techniques by providing nanoscale structural imaging with simultaneous chemical and functional analysis. However, analysis and interpretation of the collected data is complicated by the data multidimensionality and size. Here, I will discuss data analytics techniques for data co-registration and semi-automated interpretation based on multivariate statistical analysis.This research was conducted at the Center for Nanophase Materials Sciences, which is a DOE Office of Science User Facility.
Neural Networks for Processing of Low Signal-to-noise Data in Scanning Probe Microscopy: Nikolay Borodinov1; Sabine Neumayer1; Sergei Kalinin1; Olga Ovchinnikova1; Rama Vasudevan1; Stephen Jesse1; 1Oak Ridge National Laboratory
Functional fitting of the materials response is a common approach to extract physical properties. With the recent developments in scanning probe microscopy instrumentation, such response can be acquired at nanoscale level providing ability to map relevant parameters across the image and develop a deeper understanding of the systems under study. As signal-to-noise ratio decreases, the functional fits done using traditional iterative methods become very sensitive to initial guesses and yield spurious results. Here, we demonstrate a neural network-based approach for signal processing which allows for effective extraction of physical parameters at small driving signals or when a material’s response is weak.
9:40 AM Invited
Recent Advances in 3D Reconstruction Based on Spherical Indexing of EBSD Data: Marc De Graef1; 1Carnegie Mellon University
3D microstructure characterization by means of serial sectioning has become a mature field in recent years. The advent of fast sectioning systems (plasma FIB, femto-second laser) make possible the rapid acquisition of large data sets, measuring in the multi-Tb range. Converting this data into a usable 3D model remains a challenging task because many important samples have been exposed to an external influence (e.g., deformation) which often negatively impacts the signal-to-noise ratio of electron back-scatter patterns (EBSPs). Traditional Hough-based indexing approaches are now being replaced by machine learning algorithms, including dictionary-based indexing, convolutional neural networks, and spherical harmonic transform indexing. In this contribution we will highlight the state-of-the-art in EBSD data analysis as applied to large scale multi-layer multi-modal data sets; in particular we will describe example of the recently introduced spherical indexing technique which may, in the near future, rival the real-time performance of Hough-based indexing.
10:20 AM Break
Automated Defect Detection in Electron Microscopy with Machine Learning: Dane Morgan1; Mingren Shen1; Wei Li2; Kevin Field3; 1University of Wisconsin, Madison; 2Google; 3Oak Ridge National Laboratory
Electron microscopy is widely used to explore defects in crystal structures, but human tracking of defects can be time consuming, error prone, and unreliable, and is not scalable to large numbers of images or real-time analysis. In this work we discuss application of machine learning approaches to find the location and geometry of different defect clusters in irradiated steels. We show that performance comparable to human analysis can be achieved with relatively small training data sets. We explore multiple deep learning methods that provide various features, e.g., fast processing for video and pixel level categorization to simplify defect dimension determination.
10:55 AM Invited
4D STEM Data Acquisition, Analytics and Functional Material Property Extraction: Debangshu Mukherjee1; Suhas Somnath1; Alex Belianinov1; Stephen Jesse1; Raymond Unocic1; 1Oak Ridge National Laboratory
Aberration corrected scanning transmission electron microscopy (STEM) is a powerful characterization tool that allows for the structural and chemical analysis of materials at length scales down to the atomic scale using high spatial resolution imaging, electron diffraction and spectroscopy. Recent advances in high speed electron detectors has opened new opportunities to explore materials functionality using an approach termed 4D STEM. Here, a sub-Å-sized electron probe is rastered through the material of interest and the scattered electrons, containing information rich, 2D diffraction patterns acquired pixel-by-pixel at every electron probe position, are collected by the detector. The 4D STEM data sets (e.g. probe position (x,y) and diffraction pattern (x’,y’), are inherently large which necessitates a big-data analytics approach to process then analyze the data to obtain meaningful information. Here we will discuss our data analytics approach with relevant examples for strain mapping of catalyst nanoparticles and electron ptychography of 2D materials.